inplace optimization

theano/sandbox/cuda/basic_ops.py

@@ -121,14 +121,7 @@ class GpuFromHost(GpuOp):
 
     check_input = False
 
-    def __eq__(self, other):
-        return type(self) == type(other)
-
-    def __hash__(self):
-        return hash(type(self))
-
-    def __str__(self):
-        return 'GpuFromHost'
+    __props__ = ()
 
     def make_node(self, x):
         if not isinstance(x.type, tensor.TensorType):
@@ -185,7 +178,7 @@ class GpuElemwise(GpuOp):
 
     """
 
-    __props__ = ("scalar_op", "inplace_pattern", )
+    __props__ = ("scalar_op", "inplace_pattern", "sync", )
 
     nin = property(lambda self: self.scalar_op.nin)
     nout = property(lambda self: self.scalar_op.nout)

theano/sandbox/cuda/opt.py

@@ -1128,11 +1128,9 @@ def local_gpu_advanced_incsubtensor1(node):
             compute_capability = device_properties(active_device_no)['major']
             if (compute_capability < 2 or y.ndim != 2 or x.ndim != 2):
 
-                gpu_op = GpuAdvancedIncSubtensor1(
-                    set_instead_of_inc=set_instead_of_inc)
+                gpu_op = tensor.AdvancedIncSubtensor1(**node.op._props_dict())
             else:
-                gpu_op = GpuAdvancedIncSubtensor1_dev20(
-                    set_instead_of_inc=set_instead_of_inc)
+                gpu_op = GPUAdvancedIncSubtensor1_dev20(**node.op._props_dict())
             return [gpu_op(as_cuda_ndarray_variable(x),
                            as_cuda_ndarray_variable(y), *coords)]
 
@@ -1191,10 +1189,7 @@ def local_gpu_incsubtensor(node):
                 # The IncSubtensor upcast to float32 y, so we do it
                 # explicitly to move it to the GPU.
                 y = y.astype('float32')
-            ret = GpuIncSubtensor(
-                incsubt.idx_list,
-                inplace=incsubt.inplace,
-                set_instead_of_inc=incsubt.set_instead_of_inc)(
+            ret = GpuIncSubtensor(**node.op._props_dict())(
                     as_cuda_ndarray_variable(x),
                     as_cuda_ndarray_variable(y),
                     *coords)
@@ -1921,7 +1916,7 @@ def local_gpu_downsample_factor_max(node):
         if (pad) != (0, 0) or node.op.mode != 'max' or stride != ws:
             return
         if (x.owner and isinstance(x.owner.op, HostFromGpu)):
-            gpu_ds = GpuDownsampleFactorMax(**node.op._props_dict())
+            gpu_ds = GpuDownsampleFactorMax(node.op.ds, node.op.ignore_border)
             return [host_from_gpu(gpu_ds(x.owner.inputs[0]))]
 
 
@@ -2683,7 +2678,7 @@ def gpu_sparse_block_outer_opt(node):
 
         inputs = _clear_host_from_gpu(node.inputs)
 
-        return [host_from_gpu(GpuSparseBlockOuter(node.op.inplace)(*inputs))]
+        return [host_from_gpu(GpuSparseBlockOuter()(*inputs))]
 
     elif isinstance(node.op, GpuFromHost) and \
             _owner_isinstance(node.inputs[0], SparseBlockOuter):
@@ -2691,7 +2686,7 @@ def gpu_sparse_block_outer_opt(node):
         meta_node = node.inputs[0].owner
         inputs = _clear_host_from_gpu(meta_node.inputs)
 
-        return [GpuSparseBlockOuter(meta_node.op.inplace)(*inputs)]
+        return [GpuSparseBlockOuter()(*inputs)]
 
 
 @local_optimizer([GpuSparseBlockGemv], inplace=True)

theano/tensor/basic.py

@@ -3513,7 +3513,7 @@ def transpose(x, axes=None):
     """
     if axes is None:
         axes = list(range((x.ndim - 1), -1, -1))
-    ret = DimShuffle(x.broadcastable, axes, inplace=False)(x)
+    ret = DimShuffle(x.broadcastable, axes)(x)
     if x.name and axes == list(range((x.ndim - 1), -1, -1)):
         ret.name = x.name + '.T'
     return ret

theano/tensor/elemwise.py

@@ -73,8 +73,7 @@ class DimShuffle(Op):
         list can either be an index or 'x'. Indices must be encoded
         as python integers, not theano symbolic integers.
     inplace : bool, optional
-        If True, the output will be a view of the input.
-        If False (default), the output will be a copy of the input.
+        If True (default), the output will be a view of the input.
 
     Note
     ----
@@ -136,7 +135,7 @@ class DimShuffle(Op):
     check_input = False
     __props__ = ("input_broadcastable", "new_order")
 
-    def __init__(self, input_broadcastable, new_order, inplace=False):
+    def __init__(self, input_broadcastable, new_order, inplace=True):
         input_broadcastable = tuple(input_broadcastable)
         self.input_broadcastable = input_broadcastable
         new_order = tuple(new_order)
@@ -568,8 +567,7 @@ second dimension
                 # TODO: use LComplete instead
                 args.append(dim_shuffle(
                     input.type.broadcastable,
-                    ['x'] * difference + list(range(length)),
-                    inplace=False)(input))
+                    ['x'] * difference + list(range(length)))(input))
         inputs = args
 
         # HERE: all the broadcast dims have the same length now
@@ -803,6 +801,7 @@ second dimension
                 res = theano.tensor.constant(numpy.asarray(r.data),
                                              dtype=r.type.dtype)
                 return DimShuffle((), ['x'] * nd, inplace=False)(res)
+
             new_r = Elemwise(node.op, {})(
                 *[transform(ipt) for ipt in node.inputs])
             return new_r

theano/tensor/opt.py

@@ -650,35 +650,6 @@ def local_lift_transpose_through_dot(node):
         return ret
 
 
-@gof.local_optimizer([DimShuffle])
-def dimshuffle_as_view(node):
-    op = node.op
-    if not isinstance(op, DimShuffle) or op.inplace:
-        return False
-    new_op = op.__class__(op.input_broadcastable, op.new_order, inplace=True)
-    v = new_op(*node.inputs)
-    copy_stack_trace(node.outputs[0], v)
-    return [v]
-
-# Step 60 is the inplace optimization stage.
-compile.optdb.register('dimshuffle_as_view',
-                       TopoOptimizer(
-                           dimshuffle_as_view,
-                           failure_callback=TopoOptimizer.warn_inplace),
-                       60,
-                       'fast_run', 'inplace')
-register_canonicalize(local_dimshuffle_lift)
-register_specialize(local_dimshuffle_lift)
-
-
-@register_canonicalize
-@gof.local_optimizer([T.DimShuffle])
-def local_dimshuffle_no_inplace_at_canonicalize(node):
-    if isinstance(node.op, T.DimShuffle) and node.op.inplace:
-        return [T.DimShuffle(node.op.input_broadcastable,
-                             node.op.new_order, inplace=False)(node.inputs[0])]
-
-
 ######################
 # Casting operations #
 ######################

theano/tensor/tests/test_opt.py

@@ -117,7 +117,7 @@ class test_dimshuffle_lift(unittest.TestCase):
         x, y, z = inputs()
         e = ds(ds(x, (1, 0)), (1, 0))
         g = FunctionGraph([x], [e])
-        self.assertTrue(str(g) == "[DimShuffle{1,0}(DimShuffle{1,0}(x))]")
+        self.assertTrue(str(g) == "[InplaceDimShuffle{1,0}(InplaceDimShuffle{1,0}(x))]")
         dimshuffle_lift.optimize(g)
         self.assertTrue(str(g) == "[x]")
         # no need to check_stack_trace as graph is supposed to be empty
@@ -126,11 +126,10 @@ class test_dimshuffle_lift(unittest.TestCase):
         x, y, z = inputs()
         e = ds(ds(x, (1, 'x', 0)), (2, 0, 'x', 1))
         g = FunctionGraph([x], [e])
-        self.assertTrue(
-            str(g) == "[DimShuffle{2,0,x,1}(DimShuffle{1,x,0}(x))]",
-            str(g))
+        self.assertTrue(str(g) == "[InplaceDimShuffle{2,0,x,1}(InplaceDimShuffle{1,x,0}(x))]",
+                        str(g))
         dimshuffle_lift.optimize(g)
-        self.assertTrue(str(g) == "[DimShuffle{0,1,x,x}(x)]", str(g))
+        self.assertTrue(str(g) == "[InplaceDimShuffle{0,1,x,x}(x)]", str(g))
         # Check stacktrace was copied over correctly after opt was applied
         self.assertTrue(check_stack_trace(g, ops_to_check='all'))
 
@@ -138,10 +137,9 @@ class test_dimshuffle_lift(unittest.TestCase):
         x, y, z = inputs()
         e = ds(ds(ds(x, (0, 'x', 1)), (2, 0, 'x', 1)), (1, 0))
         g = FunctionGraph([x], [e])
-        self.assertTrue(
-            str(g) == "[DimShuffle{1,0}(DimShuffle{2,0,x,1}"
-                      "(DimShuffle{0,x,1}(x)))]",
-            str(g))
+        self.assertTrue(str(g) == "[InplaceDimShuffle{1,0}(InplaceDimShuffle{2,0,x,1}"
+                          "(InplaceDimShuffle{0,x,1}(x)))]",
+                        str(g))
         dimshuffle_lift.optimize(g)
         self.assertTrue(str(g) == "[x]", str(g))
         # no need to check_stack_trace as graph is supposed to be empty
@@ -179,24 +177,22 @@ class test_dimshuffle_lift(unittest.TestCase):
         m = T.matrix(dtype="float64")
         out = ((v + 42) * (m + 84)).T
         g = FunctionGraph([v, m], [out])
-        init_str_g = ("[DimShuffle{1,0}(Elemwise{mul,no_inplace}"
-                      "(DimShuffle{x,0}(Elemwise{add,no_inplace}"
+        init_str_g = ("[InplaceDimShuffle{1,0}(Elemwise{mul,no_inplace}"
+                      "(InplaceDimShuffle{x,0}(Elemwise{add,no_inplace}"
                       "(<TensorType(float64, vector)>, "
-                      "DimShuffle{x}(TensorConstant{42}))), "
+                      "InplaceDimShuffle{x}(TensorConstant{42}))), "
                       "Elemwise{add,no_inplace}"
                       "(<TensorType(float64, matrix)>, "
-                      "DimShuffle{x,x}(TensorConstant{84}))))]")
-
+                      "InplaceDimShuffle{x,x}(TensorConstant{84}))))]")
         self.assertTrue(str(g) == init_str_g)
         new_out = local_dimshuffle_lift.transform(g.outputs[0].owner)[0]
         new_g = FunctionGraph(g.inputs, [new_out])
         opt_str_g = ("[Elemwise{mul,no_inplace}(Elemwise{add,no_inplace}"
-                     "(DimShuffle{0,x}(<TensorType(float64, vector)>), "
-                     "DimShuffle{x,x}(TensorConstant{42})), "
-                     "Elemwise{add,no_inplace}(DimShuffle{1,0}"
+                     "(InplaceDimShuffle{0,x}(<TensorType(float64, vector)>), "
+                     "InplaceDimShuffle{x,x}(TensorConstant{42})), "
+                     "Elemwise{add,no_inplace}(InplaceDimShuffle{1,0}"
                      "(<TensorType(float64, matrix)>), "
-                     "DimShuffle{x,x}(TensorConstant{84})))]")
-
+                     "InplaceDimShuffle{x,x}(TensorConstant{84})))]")
         self.assertTrue(str(new_g) == opt_str_g)
         # Check stacktrace was copied over correctly after opt was applied
         self.assertTrue(check_stack_trace(new_g, ops_to_check='all'))
@@ -6301,7 +6297,7 @@ class Test_lift_transpose_through_dot(unittest.TestCase):
     def test_matrix_matrix(self):
         a, b = matrices('ab')
         g = self.simple_optimize(FunctionGraph([a, b], [tensor.dot(a, b).T]))
-        sg = '[dot(DimShuffle{1,0}(b), DimShuffle{1,0}(a))]'
+        sg = '[dot(InplaceDimShuffle{1,0}(b), InplaceDimShuffle{1,0}(a))]'
         assert str(g) == sg, (str(g), sg)
         # Check stacktrace was copied over correctly after opt was applied
         self.assertTrue(check_stack_trace(g, ops_to_check='all'))
@@ -6313,7 +6309,7 @@ class Test_lift_transpose_through_dot(unittest.TestCase):
             [a, b],
             [tensor.dot(a.dimshuffle('x', 0), b).T]),
             level='stabilize')
-        sg = '[dot(DimShuffle{1,0}(b), DimShuffle{0,x}(a))]'
+        sg = '[dot(InplaceDimShuffle{1,0}(b), InplaceDimShuffle{0,x}(a))]'
         assert str(g) == sg, (str(g), sg)
         # Check stacktrace was copied over correctly after opt was applied
         self.assertTrue(check_stack_trace(g, ops_to_check='all'))
@@ -6325,7 +6321,7 @@ class Test_lift_transpose_through_dot(unittest.TestCase):
             [a, b],
             [tensor.dot(b, a.dimshuffle(0, 'x')).T]),
             level='stabilize')
-        sg = '[dot(DimShuffle{x,0}(a), DimShuffle{1,0}(b))]'
+        sg = '[dot(InplaceDimShuffle{x,0}(a), InplaceDimShuffle{1,0}(b))]'
         assert str(g) == sg, (str(g), sg)
         # Check stacktrace was copied over correctly after opt was applied
         self.assertTrue(check_stack_trace(g, ops_to_check='all'))